https://telecombcn-dl.github.io/dlmm-2017-dcu/ Deep learning technologies are at the core of the current revolution in artificial intelligence for multimedia data analysis. The convergence of big annotated data and affordable GPU hardware has allowed the training of neural networks for data analysis tasks which had been addressed until now with hand-crafted features. Architectures such as convolutional neural networks, recurrent neural networks and Q-nets for reinforcement learning have shaped a brand new scenario in signal processing. This course will cover the basic principles and applications of deep learning to computer vision problems, such as image classification, object detection or image captioning.

1. Kevin McGuinness
kevin.mcguinness@dcu.ie
Research Fellow
Insight Centre for Data Analytics
Dublin City University
DEEP
LEARNING
WORKSHOP
Dublin City University
27-28 April 2017
Transfer Learning
Day 2 Lecture 4
1

2. Semi-supervised and transfer learning
Myth: you can’t do deep learning unless you have a million labelled examples for
your problem.
Reality
● You can learn useful representations from unlabelled data
● You can transfer learned representations from a related task
● You can train on a nearby surrogate objective for which it is easy to generate
labels
2

3. Transfer learning: idea
Instead of training a deep network from scratch for your task:
● Take a network trained on a different domain for a different source task
● Adapt it for your domain and your target task
This lecture will talk about how to do this.
Variations:
● Same domain, different task
● Different domain, same task
3

4. Transfer learning: idea
Source data
E.g. ImageNet
Source model
Source labels
Target data
E.g. PASCAL
Target model
Target labels
Transfer Learned
Knowledge
Large
amount of
data/labels
Small
amount of
data/labels
4

5. Example: PASCAL VOC 2007
● Standard classification benchmark, 20 classes, ~10K images, 50% train, 50% test
● Deep networks can have many parameters (e.g. 60M in Alexnet)
● Direct training (from scratch) using only 5K training images can be problematic. Model overfits.
● How can we use deep networks in this setting?
5

6. “Off-the-shelf”
Idea: use outputs of one or more layers of a network trained on a different task as
generic feature detectors. Train a new shallow model on these features.
conv2
conv3
fc1
conv1
loss
Data and labels (e.g. ImageNet)
fc2
softmax
TRANSFER
Shallow classifier (e.g. SVM)
conv2
conv3
fc1
conv1
Target data and labels
features
6

7. Off-the-shelf features
Works surprisingly well in practice!
Surpassed or on par with state-of-the-art in several tasks
in 2014
Image classification:
● PASCAL VOC 2007
● Oxford flowers
● CUB Bird dataset
● MIT indoors
Image retrieval:
● Paris 6k
● Holidays
● UKBench
Razavian et al, CNN Features off-the-shelf: an Astounding Baseline for Recognition, CVPRW 2014 http://arxiv.org/abs/1403.6382
Oxford 102 flowers dataset
7

8. Can we do better than off the shelf features?
8

9. labels
Fine-tuning: supervised task adaptation
Train deep net on “nearby” task for which it is
easy to get labels using standard backprop
● E.g. ImageNet classification
● Pseudo classes from augmented data
● Slow feature learning, ego-motion
Cut off top layer(s) of network and replace with
supervised objective for target domain
Fine-tune network using backprop with labels for
target domain until validation loss starts to
increase
conv2
conv3
fc1
conv1
surrogate loss
surrogate data
fc2 + softmax
real labelsreal data
real loss
my_fc2 + softmax
9

10. Freeze or fine-tune?
Bottom n layers can be frozen or fine tuned.
● Frozen: not updated during backprop
● Fine-tuned: updated during backprop
Which to do depends on target task:
● Freeze: target task labels are scarce, and
we want to avoid overfitting
● Fine-tune: target task labels are more
plentiful
In general, we can set learning rates to be
different for each layer to find a tradeoff between
freezing and fine tuning
labels
conv2
conv3
fc1
conv1
loss
data
fc2 + softmax
Finetunedfrozen
LR = 0
LR > 0
10

11. How transferable are features?
Lower layers: more general features. Transfer
very well to other tasks.
Higher layers: more task specific.
Fine-tuning improves generalization when
sufficient examples are available.
Transfer learning and fine tuning often lead to
better performance than training from scratch on
the target dataset.
Even features transferred from distant tasks are
often better than random initial weights!
Yosinki et al. How transferable are features in deep neural networks. NIPS 2014. https://arxiv.org/abs/1411.1792
11

12. 12

13. Unsupervised domain adaptation
Also possible to do domain adaptation without labels in target set.
Y Ganin and V Lempitsky, Unsupervised Domain Adaptation by Backpropagation, ICML 2015 https://arxiv.org/abs/1409.7495 13

14. Unsupervised domain adaptation
Y Ganin and V Lempitsky, Unsupervised Domain Adaptation by Backpropagation, ICML 2015 https://arxiv.org/abs/1409.7495 14

15. Semi-supervised domain adaptation
When some labels are available in the target domain, then we can use these when
doing domain adaptation. I.e. combine fine tuning and unsupervised domain
adaptation.
Tzeng et al. take this a step further and try to simultaneously optimize a loss that
maximizes:
1. classification accuracy on both source and target datasets
2. domain confusion of a domain classifier
3. agreement of classifier score distributions across domains
Tzeng, Eric, et al. Simultaneous deep transfer across domains and tasks. ICCV. 2015.
15

16. Semi-supervised domain adaptation
Tzeng, Eric, et al. Simultaneous deep transfer across domains and tasks. ICCV. 2015.
16

17. Semi-supervised domain adaptation
Classifier loss
Domain confusion
loss
Soft label loss to
align classifier
scores across
domains
Tzeng, Eric, et al. Simultaneous deep transfer across domains and tasks. ICCV. 2015.
17

18. Domain confusion loss
Alternate optimization of two objectives (like adversarial training). First makes
domain classifier as good as possible. Standard binary cross entropy loss:
Second makes features as confusing as possible for the discriminator:
Tzeng, Eric, et al. Simultaneous deep transfer across domains and tasks. ICCV. 2015.
18

19. Alignment of source and target predictions
Calculate per class
average softmax
activations in source
domain
"bottle"
Tzeng, Eric, et al. Simultaneous deep transfer across domains and tasks. ICCV. 2015.
19

20. Alignment of source and target predictions
Use these as the target
distribution for target
domain.
Minimizing cross entropy
loss same as minimizing
KL divergence!
Tzeng, Eric, et al. Simultaneous deep transfer across domains and tasks. ICCV. 2015.
20

21. Semi-supervised domain adaptation
Tzeng, Eric, et al. Simultaneous deep transfer across domains and tasks. ICCV. 2015.
21

22. Summary
● Possible to train very large models on small data by using transfer learning and
domain adaptation
● Off the shelf features work very well in various domains and tasks
● Lower layers of network contain very generic features, higher layers more task
specific features
● Supervised domain adaptation via fine tuning almost always improves
performance
● Possible to do unsupervised domain adaptation by matching feature
distributions
22

23. Questions?
23

24. Additional resources
● Lluis Castrejon, “Domain adaptation and zero-shot learning”. University of
Toronto 2016.
● Hoffman, J., Guadarrama, S., Tzeng, E. S., Hu, R., Donahue, J., Girshick, R., ... & Saenko, K. (2014). LSDA: Large scale
detection through adaptation. NIPS 2014. (Slides by Xavier Giró-i-Nieto)
● Yosinski, Jason, Jeff Clune, Yoshua Bengio, and Hod Lipson. "How transferable are features in deep neural
networks?." In Advances in Neural Information Processing Systems, pp. 3320-3328. 2014.
● Shao, Ling, Fan Zhu, and Xuelong Li. "Transfer learning for visual categorization: A survey." Neural Networks and
Learning Systems, IEEE Transactions on 26, no. 5 (2015): 1019-1034.
● Chen, Tianqi, Ian Goodfellow, and Jonathon Shlens. "Net2Net: Accelerating Learning via Knowledge Transfer." ICLR
2016. [code] [Notes by Hugo Larrochelle
● Gani, Yaroslav, Evgeniya Ustinova, Hana Ajakan, Pascal Germain, Hugo Larochelle, François Laviolette, Mario
Marchand, and Victor Lempitsky. "Domain-Adversarial Training of Neural Networks." arXiv preprint arXiv:1505.07818
(2015).
24